A Scripps Research team published details of a new AI model called ECG-CLIP on September 1, 2026, in Lancet Digital Health that can detect and predict multiple heart diseases using far fewer hand-labeled examples than current tools require. The model, trained on over 1.7 million ECGs from more than 540,000 people paired with clinicians' notes, adapts to new clinical tasks with as few as a dozen confirmed cases - a capability that could prove critical for rare disease diagnosis and resource-limited settings.
How the model was built and tested
The researchers constructed ECG-CLIP as a foundation model, a type of AI that learns from diverse datasets and then performs many downstream tasks. Unlike three other ECG-trained foundation models evaluated in the study, ECG-CLIP ingested both the electrical signal data and the natural-language clinical notes written by physicians. The team then benchmarked it against two supervised baseline models, a general foundation model not trained on ECGs, and the three ECG-only foundation models.
"Our new algorithm only needs to see on the order of a dozen confirmed ECGs of a specific disease to detect that disease in the future," said senior author Giorgio Quer, an assistant professor of digital medicine at Scripps Research. "This is similar to how a clinician would learn: not from a million examples, but from understanding the general physiology behind an ECG first and then seeing a few specific cases."
Detection performance with sparse data
For disease detection, the team tested the models on acute myocardial infarction, cardiac amyloidosis, and hypertrophic cardiomyopathy using a dataset of over 800,000 ECGs. ECG-CLIP consistently outperformed the standard models on all three diseases as measured by area under the curve. It matched the next-best model's performance while using roughly 91% less hand-labeled training data on average.
The gap between ECG-CLIP and the other ECG foundation models was widest when labeled data was scarce - as low as 10 positive examples per disease - and narrowed as more labeled examples became available. The model also performed well on single-lead ECG data for detecting heart attacks, which could make it useful in clinics or emergency settings where a full 12-lead setup is unavailable. This work sits alongside other AI applications in healthcare, such as those explored in AI for Healthcare.
Predicting future events and adverse outcomes
In a second task, ECG-CLIP outperformed all other models at predicting future atrial fibrillation from 12-lead ECGs that showed normal heart rhythms. For the third task - predicting adverse health outcomes - the model produced the best estimates of 30-day survival following an emergency department visit or surgery, as well as the likelihood of developing chronic kidney disease or type II diabetes within three years.
To address the black-box problem common to AI tools, the researchers generated saliency maps that highlight which regions of the ECG signal most influenced the model's predictions. These visual cues give clinicians a clearer view of the reasoning behind a given output.
Why this matters for science and research professionals
For researchers working with small, well-curated clinical datasets, ECG-CLIP's data efficiency changes the feasibility calculation for applying AI. Instead of requiring thousands of labeled examples to reach clinical-grade performance, teams can start with a pretrained model that already encodes physiological knowledge from waveform-text pairs. The approach also points toward a practical path for integrating AI into wearable ECG devices, where continuous remote monitoring could eventually flag rare conditions that a single spot-check would miss. The broader trend of applying foundation models in scientific domains is a growing focus in AI for Science & Research.
Your membership also unlocks:
